Output of feedfoward neural network problem.

Hi everyone,

I work with the feedfoward neural network model of the toolbox neural network of matlab. I need to change the activation function the output layer by a logsig function insteed of the purelin function.

I write the following code:

    NET = feedforwardnet(nbHidden);
    NET.layers{2}.transferFcn = 'logsig'; % Change the activation function of the output layer by the 'logsig' function.
    [NET,TR] = train(NET,Xtrain,Ytrain); % Ytrain equal to zero or one

And then I apply a test with four input and the expected output is [ 1 0 0 1 ]. But I get: [ 1 0.5 0.5 1 ]. In a more complex test, I have the same problem, the output min value is 0.5 and not 0. Did I miss something in the modification of the activation function?

Remark:

NET.outputs{2}.range

give

[ 0 1 ]

Answer 1

[Pooja Sethia]

Output of the neural network is the conditional posterior probability of the input. 'logsig' function gives 0< P(i = class| x) < 1. Therefore to get the binary output you need to normalize the probablities to unity sum after training.

Answer 2

[Rémi Souriau]

I have two classes (0 or 1) for one output. I do not get how do you want normalize the output.

I finally found another topic on the same issue.